SOURCE-REVIEWED MODEL PROFILE
Kimi-K3
A native multimodal mixture-of-experts model for long-running coding, research and knowledge work, using Kimi Delta Attention and Attention Residuals.
Metadata reviewed 2026-10-24 · Original creator: Moonshot AI
What it does
- 2.8T total parameters; 104B activated
- Native multimodal architecture
- Always-on thinking in the published usage format
Local fit
Large server deployment. Not locally benchmarked or included as a one-click desktop download. Sparse activation does not make the full weight set fit a consumer GPU.
Publisher/model configuration specification. Usable context depends on deployment resources.
Evidence and limits
Source-reviewed research profile; this exact checkpoint has not been benchmarked locally.
Research profile only
License and source
Custom model license. Open weights do not imply MIT licensing or unrestricted redistribution.
Read the original model card ↗
Reviewed source revision: f831ab66814297da540d832a5235f8e904f29d06